The Trough We're Standing In - Where AI actually is in May 2026.

On Diary of a CEO this month, Uber's CEO Dara Khosrowshahi said something that should have been the headline of the week. He told Steven Bartlett that he's been present in private conversations between executives about the "sheer amount of disruption" they expect from AI and then watched those same leaders appear on CNBC and tell audiences not to expect it. He didn't push back on the framing that other CEOs are lying. He agreed and went further.

Nobody connected that moment to what else happened this month. But they should have.

Uber's COO said the company is no longer seeing proportional productivity gains from its AI spend. The CTO had already told The Information that Uber burned through its entire $3.4 billion 2026 AI coding budget in four months. Per-engineer token costs hit $500 to $2,000 a month. Usage of Claude Code went from 32% to 84% of their 5,000-engineer organisation in a single quarter. Seventy percent of code committed at Uber now originates with AI. One in ten live backend updates ships by an agent with no human in the loop.

Microsoft began revoking internal Claude Code licences on May 14th. The deadline is June 30th - end of fiscal year. Engineers are being routed back to GitHub Copilot, the cheaper, less capable tool Microsoft already owns. The reason given is cost. The reason underneath that is the same paradox Uber is living: the better the tool works, the more engineers use it, and the higher the bill grows.

These three stories are the same story.

The mainstream AI discourse right now sits in two camps. The accelerationists who think this is all working, productivity is up, the future is bright, get on board. The doomers who think superintelligence is two years away and we're all going to die. Both camps spend most of their time arguing with each other and neither of them is describing what's actually happening inside companies right now.

What's actually happening is this:

The tools are too good. Engineers love them. Adoption rates are off the charts in the companies that have rolled them out properly. The technology is doing exactly what its makers promised it would do. And the unit economics underneath that success do not work.

This is not an AI failure. This is a deployment failure dressed up as a budget conversation.

Companies bought AI the way they buy software. Flat seat licences. Predictable monthly cost. Then they discovered that the AI tools they bought aren't software in any meaningful sense, they're consumption-priced, and the consumption scales with how useful the tool actually is. Microsoft's finance team didn't see the bill coming because nobody on the buying side understood that "engineers love it" and "we can afford it" are now opposing forces.

Underneath the cost story is a deeper one. The companies cutting Claude Code aren't cutting it because it doesn't work. They're cutting it because it works so well that their internal systems - finance, procurement, headcount planning, performance management - don't know how to absorb what it's doing. The org chart wasn't designed for a world where one engineer with the right tools does the work of ten. The budget model wasn't designed for tools that get more expensive the more your team relies on them. The performance review wasn't designed to evaluate humans whose primary skill is now orchestrating non-humans.

This is what I've been calling the human layer for two years now. It's the layer that almost nobody is building.

Khosrowshahi's admission matters more than the budget numbers. Because he's saying, on a top-five global podcast, that the executives shaping public AI discourse are lying. Not exaggerating. Lying. They are telling the public that productivity will go up and society will adapt while privately preparing for displacement at a scale they're not willing to name. That dissonance is not a communication problem. It's a trust problem. And it is going to define the next phase of AI adoption inside companies more than any model release or token-pricing shift.

Here's what happens when leadership says one thing publicly and another thing privately: the smart people in the building figure it out. They always do. The "secret power users" already automating half their job aren't telling anyone because they don't trust that the upside of disclosing will outweigh the downside. The "quiet resisters" aren't touching the tools because they've correctly read that the company doesn't really believe its own optimism. The "anxious self-protectors" are right to be anxious. The leadership team calling all of this "slow adoption" is the part of the system that's most out of touch with what's actually happening underneath them.

You can't fix that with another rollout. You can't fix it with training. You can't fix it by buying a better tool or switching to a flat-rate provider. You fix it by telling the truth, and then designing the human system to absorb what the truth implies.

So here is where we actually are in May 2026.

The technology is working. The economics are breaking. The leadership is not being honest about what comes next. The humans inside the system know all three things even when nobody says them out loud. Gartner has now placed generative AI in the trough of disillusionment, predicting that 25% of planned 2026 AI budget will slip into 2027 as proofs of concept die quietly in procurement pipelines that don't know what they're looking at. The data confirms what the stories already told us.

And the companies that will come out of the next twelve months in good shape are the ones that stop pretending it's a tooling problem and start treating it as what it actually is.

A trust problem. A communications problem. A people problem.

A system that was designed for a previous version of work, trying to absorb something the previous version was never built for.

Uber didn't blow $3.4 billion in four months because Claude Code was bad. They blew it because nobody at the boundary between engineering and finance was thinking about the shape of the system that was forming underneath them. Microsoft didn't pull Claude Code because it failed. They pulled it because it worked so well it exposed a structural problem they had no language for.

Khosrowshahi isn't telling us the future. He's describing the present. The dissonance is here. The deployment problem is here. The human layer is here, breaking quietly, in the gap between what leadership says and what the people in the building can see for themselves.

This is the work.

Not the discourse. Not the takes. Not the predictions about 2027.

This. Right here. Now. The system that has to learn to hold.

Gail Weiner

AI Trust Architect. Bristol, May 2026.